Quantifying plastic pollution on surface water is essential to understand and mitigate the impact of plastic pollution to the environment. Current monitoring methods such as visual counting are labor intensive. This limits the feasibility of scaling to long‐term monitoring at multiple locations. We present an automated method for monitoring plastic pollution that overcomes this limitation. Floating macroplastics are detected from images of the water surface using deep learning. We perform an experimental evaluation of our method using images from bridge‐mounted cameras at five different river locations across Jakarta, Indonesia. The four main results of the experimental evaluation are as follows. First, we realize a method that obtains a reliable estimate of plastic density (68.7% precision). Our monitoring method successfully distinguishes plastics from environmental elements, such as water surface reflection and organic waste. Second, when trained on one location, the method generalizes well to new locations with relatively similar conditions without retraining (≈50% average precision). Third, generalization to new locations with considerably different conditions can be boosted by retraining on only 50 objects of the new location (improving precision from ≈20% to ≈42%). Fourth, our method matches visual counting methods and detects ≈35% more plastics, even more so during periods of plastic transport rates of above 10 items per meter per minute. Taken together, these results demonstrate that our method is a promising way of monitoring plastic pollution. By extending the variety of the data set the monitoring method can be readily applied at a larger scale.

中文翻译：

---

使用深度学习和摄像头的自动化河道塑料监测

量化地表水上的塑料污染对于理解和减轻塑料污染对环境的影响至关重要。当前的监视方法（例如视觉计数）是劳动密集型的。这限制了扩展到多个位置进行长期监视的可行性。我们提出了一种自动方法来监测塑料污染，克服了这一限制。使用深度学习从水面图像中检测出漂浮的大塑料。我们使用来自印度尼西亚雅加达五个不同河流位置的桥式摄像机的图像对方法进行了实验评估。实验评估的四个主要结果如下。首先，我们实现了一种能够获得可靠的塑性密度估计值（精度为68.7％）的方法。我们的监控方法成功地将塑料与环境元素（例如水表面反射和有机废物）区分开。其次，在一个位置上进行训练时，该方法可以很好地推广到条件相对相似的新位置，而无需重新训练（≈50％的平均精度）。第三，可以通过仅对新位置的50个对象进行再训练来提高对条件相当不同的新位置的泛化能力（将精度从≈20％提高到≈42％）。第四，我们的方法与视觉计数方法相匹配，可以检测到≈35％的塑料，甚至在每分钟每米10件以上的塑料运输速度期间也能检测到更多的塑料。综上所述，这些结果表明我们的方法是监测塑料污染的一种有前途的方法。通过扩展数据集的多样性，可以轻松地大规模应用监视方法。